Looking Under the Hood: Live Digital Twin Supports Operational Decisions for Stable Biological Phosphorus Removal

INTRODUCTION Biological Phosphorus Removal (BioP) offers advantages such as reduced metal salt use, lower chemical sludge solids, and improved settleability. However, its complex biochemical relationships are sensitive to operational changes and prone to upsets that threaten effluent quality. Operators at the City of Ann Arbor Water Resource Recovery Facility (WRRF) rely on lab data, SCADA, and qualitative observations to optimize performance. In 2025, the City partnered with Jacobs Engineering to implement a Hybrid Optimizer (a live digital twin) for insight into BioP upsets and to provide operational decision support. To the authors' knowledge, this is the first plant-wide live hybrid digital twin for BioP. The Ann Arbor WRRF is a 29.5 MGD (112 MLD) advanced secondary treatment facility (recent annual average 13.7 MGD). Liquids treatment includes headworks, primary clarification, anoxic/anaerobic/aerobic bioreactors, final clarification, filtration, and UV disinfection before discharge to the Huron River. Solids treatment includes lime addition for phosphorus precipitation to mixed PS and TWAS before centrifuging. The Hybrid Optimizer combines machine learning and mechanistic modeling for full-scale nutrient management, developed under WRF Project 5121 (Johnson et al., 2024). Following project completion, the tool has undergone continuous refinement and has been applied at additional facilities, including Ann Arbor. This paper summarizes the facility's interaction with a live digital twin and highlights key operational insights from the Hybrid Optimizer deployment at Ann Arbor. METHODOLOGY Deployment steps mirrored previous live applications, outlined in the WRF 5121 Final Report and in previous publications (Yang et al. 2024, Yang et al. 2023; Registe et al. 2023; Menniti et al. 2023; Johnson et al. 2023; Oristian et al. 2023). The critical role of the facility is highlighted here. [b]Data Ingestion & Preparation:[/b] Continuous collaboration with staff ensured robust data pipelines, accurate tag interpretation, and reliable metadata. [b]Model Development:[/b] Seasonal load changes from the local university require frequent changes of basins in service. As basins often exhibit different performance, this required modeling individual trains in Sumo24©. Staff input was critical for dynamic solids operations and translating facility variability into the mechanistic model. [b]Machine Learning & Recommendations:[/b] Staff prioritized KPIs such as primary sludge blanket age and secondary loading when selecting information to daylight from the digital twin. Staff identified daily emails as the most useful way to receive insights into these KPIs and what operational changes are recommended as a result (see Figure 1 for example email). Influent forecasting is in progress. [b]User Interface:[/b] Jacobs and staff designed a GUI to display key soft sensor results and KPIs (Figure 3). Staff also opted for downloaded soft sensor data into a Sumo© model for offline 'what-if' analysis, saving many hours of manual input and enabling high-frequency simulation of costly-to-measure parameters. RESULTS AND DISCUSSION [b]Soft Sensor Predicts Process Upsets:[/b] The hybrid optimizer employed autocalibration of key parameters such as alpha_sat and hydrolysis to track the influent strength and effluent phosphorus performance long term. The system was able to predict influent laboratory data and effectively anticipate both biological phosphorus removal (BioP) and nitrification upsets, as shown in dashboard visualizations during these events (Figure 2). The model predicts slightly earlier and more extreme upsets, which is preferable as a warning signal for operators. [b]User Interface Daylights Unmeasurable BioP Indicators:[/b] The Hybrid Optimizer tracks key indicators such as PHA uptake, VFA-to-phosphorus feed ratio, and GAO/PAO balance to assess BioP stability. These mechanistic and soft-sensor insights reveal how primary fermentation and hydrolysis drive VFA production, which sustains PAO activity and prevent GAO dominance. Upsets in the simulated and measured secondary effluent OP spiked beyond the same threshold for these indicators, implying the model is capturing real BioP health (Figure 3). Staff have observed qualitative indicators of fermentation like primary clarifier sheen correlate with target indicator ranges. Unlike traditional monitoring, these predictive indicators enable proactive control of primary sludge blanket SRT and upset avoidance. Operational recommendations began in December 2025; secondary SRT guidance is under development. CONCLUSION Mechanistic and machine-learning calibration of hydrolysis rates to match BioP performance successfully produced the first live digital twin of its kind for biological phosphorus removal. The Ann Arbor Hybrid Optimizer pilot transforms advanced biokinetic modeling into practical, actionable insights for WRRF staff by leveraging a complex model of six independent trains to capture loading and wasting variability. Soft sensors provided critical data that cannot be measured directly, while developed KPIs and primary wasting recommendations were validated against both operator experience and real-time BioP upset events. The Ann Arbor Hybrid Optimizer pilot is ongoing; additional results will be provided in final paper.